Improving Ptz Camera Control for Surveillance Using Proximal Policy Optimization

Authors:
Publisher website
Google Scholar
Abstract
This paper presents advancements in camera-only surveillance systems, focusing on the inte-gration of Proximal Policy Optimization (PPO) and its recurrent variant to improve the detec-tion and tracking capabilities of PTZ cameras. Building on previous research, we refine our PPO model by adjusting observation representations and hyperparameters, significantly improving target coverage and tracking accuracy. The study evaluates two simulated scenarios, an abstract grid-like environment and a more realistic scenario with predefined intruder paths. The results demonstrate that the standard PPO outperforms both the heuristic and recurrent models. The findings confirm the potential of Reinforcement Learning (RL) in improving autonomous sur-veillance systems but also point to the need for more complex simulation environments to fully exploit the capabilities of recurrent neural networks.